For inclusion in the upcoming Linux kernel 6.13 branch, a patch has been proposed with a reworked implementation of the CRC32C checksum algorithm. The code for the CRC32C implementation has been reduced by about 10 times (from 4546 to 418 bytes). With retpoline protection against Spectre-class attacks disabled, the performance gain using the new implementation reaches 11.8% on AMD Zen 2 processors, 6.4% on Intel Emerald Rapids, and 4.8% on Intel Haswell. When retpoline is enabled, the performance gain is more pronounced, reaching 66.8% on systems with Intel Emerald Rapids processors, 35.0% on Intel Haswell, and 29.5% on AMD Zen 2. retpoline enabled | 512 | 833 | 1024 | 2000 | 3173 | 4096 | ———————+——-+——-+——-+—— +——-+——-+ Intel Haswell | 35.0% | 20.7% | 17.8% | 9.7% | -0.2% | 4.4% | Intel Emerald Rapids | 66.8% | 45.2% | 36.3% | 19.3% | 0.0% | 5.4% | AMD Zen 2 | 29.5% | 17.2% | 13.5% | 8.6% | -0.5% | 2.8% | retpoline disabled: | 512 | 833 | 1024 | 2000 | 3173 | 4096 | ———————+——-+——-+——-+—— +——-+——-+ Intel Haswell | 3.3% | 4.8% | 4.5% | 0.9% | -2.9% | 0.3% | Intel Emerald Rapids | 7.5% | 6.4% | 5.2% | 2.3% | -0.0% | 0.6% | AMD Zen 2 | 11.8% | 1.4% | 0.2% | 1.3% | -0.9% | -0.2%
The original version of CRC32C included 128 unrolled loops, which led to quite a large code size. Since modern processors with out-of-order execution support can execute commands in parallel, such optimization of branch instructions in loops has become excessive and resulted only in unnecessarily large code. Instead of 128 iterations, the new version includes only 4, significantly reducing code size and accelerating the operation.
Kilde: opennet.ru
